references/reconciliation.md

# Reconciliation and data quality

A validation query should express a falsifiable invariant. Pair counts with business-measure totals, per-key shape and date coverage. At a known grain:
- Uniqueness test: group by expected key; any COUNT(*) > 1 is a grain failure.
- Non-null critical keys: explicit IS NULL count.
- Match coverage: left anti-join or equivalent.
- Financial totals: source-vs-target differences using agreed precision, currency and temporal rules.
- Freshness: max committed event time, not merely ingestion heartbeat.
- Late data: compare event and arrival timestamps and known SLA windows.

Define failure action: block publication for severe key duplicates or missing facts, quarantine suspect records where appropriate, and warn only for tolerable thresholds. Note how known business exceptions are documented and monitored.

Never claim checks passed unless you have actual output. Provide SQL probes, expected criteria and an evidence table with `not run` state when appropriate.